GNOSIS: an R Shiny app supporting cancer genomics survival analysis with cBioPortal

Implementation

Overview. GNOSIS was initially developed to enable the exploration, visualisation and analysis of the METABRIC clinical and CNA summary data obtained from cBioPortal, as detailed in King et al. (2021)¹⁶, and the following description of its operational capabilities have their basis in that study. Although GNOSIS accepts multiple file types, including comma-, semicolon- or tab-delimited, the default settings are suited to files downloaded from cBioPortal. If users wish to upload clinical or summary genomics data files from other sources, care should be taken to set appropriate default values. GNOSIS leverages a number of R packages, primarily shiny, tidyverse, ggplot2, survival, survminer, rpart, partykit and maftools^18–26. A full list is provided in the Operation subsection. It allows users to carry out a comprehensive visual exploration and statistically robust survival analysis in a fast, simple and reproducible way, and we illustrate GNOSIS functionalities by referring to example screenshots of its operation at various relevant steps of the analysis documented in King et al. (2021)¹⁶.

Data upload and formatting. In Figure 1 we show the GNOSIS front-end with the specific entry points and ‘tabs’ highlighted. GNOSIS accesses files locally on the user’s file system, and in its default configuration, is optimised to use data files downloaded from cBioPortal. In the Input Files tab, users are provided with a space to upload the clinical patient and sample data, summary CNA data and mutation data. Whilst we have configured GNOSIS to work exclusively with both CNA and mutational data files, modification of the codebase allows users to reconfigure the GNOSIS GUI to import other genomic tracks from cBioPortal, as required. A preview of the data is provided in the GNOSIS viewing panel to ensure that the data has been read in correctly. It should be noted that the clinical patient and sample data should contain a column named "PATIENT_ID" and the CNA data should contain a column called "Hugo_Symbol". As these are core named data types for all subsequent analytics, warnings will be produced and downstream analysis will not be possible if they are missing.

Once the data is uploaded, further exploration of specific columns can be done using the Exploratory Tables tab, where up to five columns can be selected in the box sidebar and viewed. The columns should be selected in sequential order; if this is not adhered to an error will be displayed.

Figure 1. GNOSIS GUI with highlighted interface elements.

(1) The Exploratory Tables tab is selected in the tab sidebar. (2) Within tab panels allowing multiple operations to be carried out and viewed in the one tab. (3) Box sidebar allowing users to select inputs, alter arguments and customise and export visualisations. (4) Viewing panel displaying output.

Before more extensive data exploration and analysis, users are encouraged to carry out pre-processing to ensure data is in the desired format using the Recode/Subset Data tab. Users are provided with a workspace to view the variables present in the data, their type and factor levels. Users can change variables to numeric or factors using the box sidebar, which contains a space to select relevant variables.

Subsequently, users can subset the data based on up to three categorical variables and carry out survival variable recoding. In cases where CNA data is uploaded, users may produce and segment CNA metrics for each patient, as well as select and extract specific genes for further analysis. After each operation, the space to explore variable information is updated. This allows users to confirm their alterations have been implemented correctly. These operations ensure that the data is in the correct format for downstream analysis.

To allow users to save their formatted file, a space within the tab is provided. If this exported data is uploaded to GNOSIS, formatting of categorical data may have to be carried out again due to the default stringsAsFactors argument implemented when uploading data in R. In Figure 2 (A), we show how a given data file can be examined and filters applied to extract a subset, and in Figure 2 (B), the resulting subset following calculation of CNA metrics, including absolute CNA score, amplification score and deletion score for each patient, and subsequent quartile segmentation of the CNA scores is shown.

Data visualisation. Within the Exploratory Plots tab, GNOSIS provides users with a range of visualisations including boxplots, scatterplots, barplots, histograms and density plots. For example, in Figure 2 (C), a set of patient genome absolute CNA scores can be segmented - and labelled as such - into multiple equally-sized groups. These visualisations are implemented using the R package ggplot2²⁰. This way, clinical and genomic data can be interrogated and visualised separately or in combination. For each visualisation, users can use the box sidebar to select which columns to interrogate, choose whether to include NA values in the plots, choose whether to display a legend and change the plot title, x- and y-axis titles and legend titles, among others. Further options available to users include the ability to produce boxplots where the sample size is reflected in the width of the boxplot, produce scatterplots coloured by an additional variable, and to produce plain, segmented and faceted histograms and density plots. Users can also download all the resulting plots as .pngs or .svgs in specified dimensions.

Figure 2.

(A) The Recode/Subset tab, where data is being subsetted based on subtype, luminal A and luminal B subtypes are selected. (B) The dataset after CNA metrics have been calculated and quartile segmented. (C) A density plot of the resulting quartile segmentation.

Statistical and survival analysis. The primary function offered by GNOSIS is statistically robust survival analysis. GNOSIS contains several step-wise tabs to provide a complete survival analysis of the data under investigation.

Initially the Kaplan-Meier (KM) Plots tab provides survival plots and the corresponding logrank tests to identify survival-associated categorical variables, both visually and statistically (Figure 3 (A)). The Kaplan-Meier Plots tab contains three sub-tabs which provide users with spaces to produce KM plots and logrank tests for selected clinical variables, for segmented CNA variables and for variables of interest split on treatment assignment (i.e. where patients received different treatments, e.g. split into patients who received radiotherapy and patients who did not). Within each sub-tab, the interface allows users to indicate which columns contain the survival time, event status (Overall Survival (OS) or Disease Specific Survival (DSS)) and the variable of interest. It should be noted that when producing KM curves for variables split by treatment assignment the selected treatment variable must be coded as a binary YES/NO. These KM curves can be customised and exported as .pngs or .svgs using the sidebar options.

Figure 3.

(A) Kaplan-Meier (KM) plot for luminal breast cancer disease specific survival (DSS) for each CNA quartile group. The p-value associated with the logrank test and a risk table displaying the number of patients at risk at each time interval is displayed. (B) Example of a χ² analysis of the data, individual χ² tests displayed in top box and table with adjusted p-values displayed in bottom box. (C) Example of an implementation of a multivariable Cox model.

The Association Tests tab uses association tests to identify variables that are linked to each other and enables users to identify potential confounding variables in the analysis. Variable selection is done within the box sidebar and statistical association tests available include the χ² test, Fisher’s exact test, simulated Fisher’s exact test, ANOVA, Kruskal-Wallis test, pairwise t-test and Dunn’s test. The χ² test is used to assess the association between two categorical variables with sufficient cell sizes in the two-way table of categorical variables (Figure 3 (B)). Fisher’s exact test can be used in the case where any cell size is sufficiently small. ANOVA can be used to test whether there is a difference in means between groups and the Kruskal-Wallis test may be used in the situation where the assumptions of the ANOVA test are violated. Pairwise comparisons using t-tests and Dunn’s test are also available. In all cases, results of each individual association test are displayed alongside the adjusted p-values calculated using the Benjamini-Hochberg p-value adjustment. Users also have the option to carry out a basic descriptive analysis by groups using the compareGroups package²⁷. It is important that users make sure they run the appropriate statistical tests for the question of interest, that all relevant assumptions are met and that the output is interpreted correctly. To aid users in this, information buttons containing links to useful resources are available throughout the app.

In the Cox Proportional Hazards (PH) models tab, users are provided with a workspace to produce both univariate and multivariable Cox models to identify survival-associated variables, and test the assumptions of these models using graphical diagnostics based on the scaled Schoenfeld residuals (Figure 3 (C)). The Cox PH model is a regression model commonly used to investigate the association between the survival time of patients and predictor variables. The Cox PH model works for both continuous and categorical variables and extends survival analysis methods to simultaneously assess the effect of several risk factors on survival time. To produce the univariate and multivariable Cox models, the box sidebar enables the selection of the columns that contain the survival time, event status and the variables to be included in the models. The output of each univariate Cox model is displayed along with a summary table containing the adjusted p-values calculated using the Benjamini-Hochberg p-value adjustment. The validity of the PH assumption of each multivariable model fitted can be assessed by producing visualisations based on the scaled Schoenfeld residuals. The Schoenfeld residuals are independent of time, and therefore a plot displaying a non-random pattern against time indicates that the PH assumption may be violated. Where a non-significant relationship between residuals and time is observed, the PH assumption is met. Again, these plots can be customised and exported in portable network graphics (PNG) format or scalable vector graphics (SVG) format.

Following multivariable Cox model selection, users are given the option to produce corresponding adjusted survival curves, which are survival curves adjusted for the covariates in the multivariable Cox model. Within the Adjusted Survival Curves tab, users are provided with a workspace to view the multivariable Cox model that was fitted in the previous tab. This will aid users when creating the new data frame needed to produce the adjusted survival curves. Users are provided with a space to set up the new data frame including the grouping variable, variable of interest and the variables to be kept constant (Figure 4 (A)). It should be noted that all variables included in the multivariable Cox model should be included in the new data frame. Plots displaying all the adjusted survival curves and adjusted survival curves split based on grouping variable level are displayed for users to view, customise and download in PNG or SVG format.

Figure 4.

(A) Output showing format of new data and adjusted survival curves corresponding to the previous multivariable Cox model. (B) Example output of a ctree survival tree analysis. (C) Sample output from use of the maftools package, a MafSummary plot is displayed.

In the case where the PH assumption of the multivariable Cox model is violated, users can apply recursive partitioning survival trees available in the Survival Trees tab. Users can use the rpart or ctree^23–25 algorithms with customised argument parameters to produce survival trees containing one or more variables along with the corresponding KM curves (Figure 4 (B)). Users are provided with a workspace to select the survival time, event status and variables of interest, and information on each of the customisable arguments is given by pressing the information button located at the top of the box. Similar to previous tabs, the survival trees and accompanying KM curves can be exported with specified plot width and height. It should be noted that the ctree algorithm will only work where the selected categorical variables are in factor form.

Mutation analysis. An additional function of GNOSIS is the ability to summarise, analyse and visualise mutation annotation format (MAF) files using maftools²⁶. MAF files are used to store detected somatic variants and are usually provided as part of the cBioPortal downloads. The Maftools Summary tab in GNOSIS allows users to view the MAF summary, sample summary, gene summary and summary of the associated clinical data, if available. If clinical data are provided users need to make sure the column named "Tumor_Sample_Barcode" is present. These summaries provide a basic view of the uploaded MAF file and contain information on the number of mutations, type of mutations and genes affected by these mutations. The Maftools Summary tab also enables users to examine the mutational landscape of the tumours in a graphical way. The plots available include MAF summary plots which display the number of variants in each sample as a stacked barplot and variant types as a boxplot summarized by variant classification (Figure 4 (C)). GNOSIS also contains panels for oncoplots, oncostrips, graphs displaying transition and transversion rates, lollipop plots for up to three genes simultaneously, mutation load plots and somatic interaction plots, all derived from the original maftools package. Other functions of this tab include allowing users to customise and export these plots in PNG or SVG format with specific dimensions.

Reproducible research

GNOSIS facilitates reproducible research by providing a Download Code/Log tab where users can view and download a log containing information on all the inputs selected throughout the session, as well as downloading an R script containing code to reproduce the outputs displayed in the app (Figure 5).

Figure 5.

Dataframe containing log of inputs selected, which can be downloaded as a .txt file. Option to download R script containing code run in app also available.

Operation

GNOSIS is available on shinyapps.io and GitHub. This enables users to access GNOSIS via a web browser or run GNOSIS locally by downloading, extracting and launching the app manually in RStudio, or running the app in RStudio using: shiny::runGitHub(repo=‘GNOSIS’,username = ‘Lydia-King’,ref="main"). The latter is recommended due to resource limitations imposed by the operators of the shinyapp.io website.

GNOSIS was originally developed using R version 3.5.1 but due to subsequent package updates, it now works on R versions _≥ 4.0.0. GNOSIS depends on the packages, BiocManager, shiny, shinymeta, shinydashboard, dashboardthemes, shinydashboardPlus, shinyWidgets, shinycssloaders, shinylogs, fontawesome, DT, tidyverse, ggplot2, fabricatr, reshape2, operator.tools, rpart, rpart.plot, partykit, coin, survminer, survival, stats, rstatix, DescTools, car, compareGroups, R.utils, RColorBrewer and maftools^18–48, which are automatically installed and loaded when running GNOSIS manually from RStudio. Should GNOSIS be run using the runGitHub() function, shiny must be installed beforehand.